AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

The question is how much of an uplift can AMD get from a big Navi/Vega 7nm chip?

HOCP just did a series of articles comparing various NVIDIA and AMD GPUs across 15 games, the transition from FuryX with it's limited 4GB RAM and 28nm to Vega 64 with it's 8GB RAM and 16nm only resulted in a 30% uplift. Compare that to a 1080Ti to 980Ti and the uplift is 70%!

It gets worse as you go down in generations, the FuryX is only 20% faster than 390X. While a 980Ti is 46% faster than 780Ti.

Granted NVIDIA's uplift 1080Ti to 2080Ti slowed down, but that's because of 16nm to 12nm transition. We know nothing about their 7nm uplift. And If AMD can repeat a similar 30% uplift with Navi/Vega 7nm this puts them on par with 1080Ti. Leaving the 2080Ti and the 7nm NVIDIA flagship untouchable. AMD badly needs to increase their uplift this round.

Thoughts?

https://www.hardocp.com/article/2018/09/04/amd_gpu_generational_performance_part_1/1
https://www.hardocp.com/article/2018/08/07/nvidia_gpu_generational_performance_part_2/17

For what it's worth, I'd expect Navi 10 to have no more CUs than Vega 10, but higher efficiency (i.e. more perf/CU/MHz, but also less power from micro-architectural optimisation), maybe around +10~20%, maybe a bit more, since there's some reason to think that Vega may be a bit buggy, and/or rough around the edges. Then 7nm should bring something like +15~25% on the clockspeed front, so a total of +27~50%. It's hard to say that this is anything more than a guess, though.

All in all that should put it somewhere between the (1080Ti/)2080 and the 2080 Ti, both in terms of performance and, presumably, power. Clearly, that wouldn't be great, but it would be competitive against Pascal/Turing. I have a feeling that AMD will have a few months of 7nm monopoly before NVIDIA responds, so that's something. After that, they'll need real progress on the micro-architecture front, which doesn't seem to be in the pipes for Navi, but might be for whatever comes next.
 
Could you elaborate more?

In case of 390 vs 580 the difference is clocks, right? 1GHz for Hawaii vs 1.4GHz for Polaris.

Clocks are part of it. However there are other aspects, e.g. Polaris has to do with a lot smaller memory bandwidth.
 
Could you elaborate more?

In case of 390 vs 580 the difference is clocks, right? 1GHz for Hawaii vs 1.4GHz for Polaris.
1.4 GHz for Polaris? It's 1266 MHz boost for Radeon RX 480 and the card reaches barely 1200 MHz in gaming workloads. It still performs at the average like Radeon R9 290X in uber mode. 20 % higher clocks don't explain performance of Radeon R9 290X with halved memory bus and halved number of ROPs. It just compensates 20 % less CUs.
 
Sure, here comes the GCN memory compression and other bunch of minor improvements. It helps a great deal to b/w constrained SKUs. The article and surrounding discussion was all about highend SKUs and their absolute scaling. That's where the rough edges show themselves.
 
Sure, here comes the GCN memory compression and other bunch of minor improvements. It helps a great deal to b/w constrained SKUs. The article and surrounding discussion was all about highend SKUs and their absolute scaling. That's where the rough edges show themselves.

And even that article concludes contrary to your statement of no improvements.
 
1.4 GHz for Polaris? It's 1266 MHz boost for Radeon RX 480 and the card reaches barely 1200 MHz in gaming workloads. It still performs at the average like Radeon R9 290X in uber mode. 20 % higher clocks don't explain performance of Radeon R9 290X with halved memory bus and halved number of ROPs. It just compensates 20 % less CUs.
You should mention though, that R9 290X ref card even in Uber mode did not always keep it's (advertised as "up to", so no blame here) clock speeds, but went down towards the 900 MHz mark. edit: Not quite, it turns out. See here. From there, 1200 MHz on Polaris is 33% already. Memory bus and ROPs were halved between Hawaii and Polaris yes, but memory speed went up 60 % IIRC (1250 to 2000), L2 cache was doubled (1 to 2 MB, per partition it was even quadrupled), DCC was introduced which all helped performance not to tank as much at it would have otherwise. Then, there's the 4 GB vs 8 GB advantage but that's mostly not changing the underlying architecture, which has served AMD quite well for a long time now.
 
Last edited:
Honestly I think Vega is broken in hardware somewhere. I doubt this is all it was supposed to be. The comment from a Linux driver dev that Primitive Shaders won't be supported for GFX9 but will for GFX10 makes me believe something is broken.
 
Honestly I think Vega is broken in hardware somewhere. I doubt this is all it was supposed to be. The comment from a Linux driver dev that Primitive Shaders won't be supported for GFX9 but will for GFX10 makes me believe something is broken.

It's hard to say whether that's due to a hardware bug, or just the fact that the hardware turned out to be harder to program than expected. But either way, something didn't go as planned, I agree that this much seems clear.
 
It's hard to say whether that's due to a hardware bug, or just the fact that the hardware turned out to be harder to program than expected. But either way, something didn't go as planned, I agree that this much seems clear.
Yeah, I'm looking forward to any Vega post mortem we might get, if someone manages to dig out enough details.
 
It'd be a very nice read about Vega and Bulldozer failures. IIRC, the R600 story was partially revealed in an interview. However, it doesn't seem to be good PR-wise to wash this kind of dirty laundry in public...

Success stories are a different matter - Zen architecture or Threadripper SKU background stories.
 
You should mention though, that R9 290X ref card even in Uber mode did not always keep it's (advertised as "up to", so no blame here) clock speeds, but went down towards the 900 MHz mark. From there, 1200 MHz on Polaris is 33% already. Memory bus and ROPs were halved between Hawaii and Polaris yes, but memory speed went up 60 % IIRC (1250 to 2000), L2 cache was doubled (1 to 2 MB, per partition it was even quadrupled), DCC was introduced which all helped performance not to tank as much at it would have otherwise. Then, there's the 4 GB vs 8 GB advantage but that's mostly not changing the underlying architecture, which has served AMD quite well for a long time now.
I'm sorry, but I don't remember, that R9 290X in Uber mode ran at 900 MHz. It was ~850 MHz in standard mode and 1 GHz in Uber with reference board:
https://www.computerbase.de/2013-10/amd-radeon-r9-290x-test/12/#abschnitt_powertune_20_in_der_praxis
 
Yep, remember this? https://www.anandtech.com/show/2679

Stories where things go wrong are no less interesting. (Poke @Ryan Smith)
It's certainly something that would be fun to do. Whether AMD is interested in doing it will be another matter. No one talks as much now as they used to, unfortunately. So the best we could hope for is finally opening up once Navi is ready.
 
It's certainly something that would be fun to do. Whether AMD is interested in doing it will be another matter. No one talks as much now as they used to, unfortunately. So the best we could hope for is finally opening up once Navi is ready.

I would imagine that it would require Navi to be pretty stellar to get on-the-record cooperation from AMD because usually company officials would only be interested in discussing past failures as a “lessons learned/what went wrong” postmortem subplot within the main redemption arch of highly successful new architecture launch.
 
It's certainly something that would be fun to do. Whether AMD is interested in doing it will be another matter. No one talks as much now as they used to, unfortunately. So the best we could hope for is finally opening up once Navi is ready.

I would imagine that it would require Navi to be pretty stellar to get on-the-record cooperation from AMD because usually company officials would only be interested in discussing past failures as a “lessons learned/what went wrong” postmortem subplot within the main redemption arch of highly successful new architecture launch.

Yup, and even then, it's unlikely that they would agree to anything as long as Vega is selling in any shape, in any kind of product, which is going to be the case for a few years at least. But at some point, it might be an option. And who knows, maybe a few years from now, if AMD remains unwilling to talk, Raja Koduri might be more open, provided he's contractually allowed to say anything.
 
If that's truly the same thing, they can put an "*" saying "* it's working for us".
Based on the whitepaper it's something devs need to implement, not automatic, and isn't that the current state on AMD side too?
 
Back
Top